Home
last modified time | relevance | path

Searched refs:vk22x0123 (Results 1 – 22 of 22) sorted by relevance

/external/XNNPACK/src/f32-dwconv/gen/
Dup4x25-minmax-wasmsimd-x86.c303 const v128_t vk22x0123 = wasm_v128_load(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
304 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
420 const v128_t vk22x0123 = wasm_v128_load(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86() local
421 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86()
Dup4x25-minmax-wasmsimd-x86-acc2.c303 const v128_t vk22x0123 = wasm_v128_load(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
304 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
422 const v128_t vk22x0123 = wasm_v128_load(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2() local
423 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_x86_acc2()
Dup4x25-minmax-sse.c303 const __m128 vk22x0123 = _mm_load_ps(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
304 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
419 const __m128 vk22x0123 = _mm_load_ps(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse() local
420 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse()
Dup4x25-minmax-sse-acc2.c303 const __m128 vk22x0123 = _mm_load_ps(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
304 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
421 const __m128 vk22x0123 = _mm_load_ps(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2() local
422 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__sse_acc2()
Dup4x25-wasmsimd.c301 const v128_t vk22x0123 = wasm_v128_load(w + 92); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
302 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
416 const v128_t vk22x0123 = wasm_v128_load(w + 92); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd() local
417 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_ukernel_up4x25__wasmsimd()
Dup4x25-minmax-neonfma.c258 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
259 vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
368 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma() local
369 vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma()
Dup4x25-minmax-wasmsimd-arm.c303 const v128_t vk22x0123 = wasm_v128_load(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
304 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
420 const v128_t vk22x0123 = wasm_v128_load(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm() local
421 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm()
Dup4x25-minmax-neonfma-acc2.c258 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
259 vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
370 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2() local
371 vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neonfma_acc2()
Dup4x25-minmax-wasmsimd-arm-acc2.c303 const v128_t vk22x0123 = wasm_v128_load(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
304 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
422 const v128_t vk22x0123 = wasm_v128_load(w + 92); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2() local
423 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up4x25__wasmsimd_arm_acc2()
Dup4x25-minmax-neon-acc2.c258 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
259 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
370 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2() local
371 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon_acc2()
Dup4x25-minmax-neon.c258 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
259 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
368 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up4x25__neon() local
369 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up4x25__neon()
Dup8x25-minmax-sse.c371 const __m128 vk22x0123 = _mm_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
373 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
544 const __m128 vk22x0123 = _mm_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
545 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
660 const __m128 vk22x0123 = _mm_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse() local
661 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse()
Dup8x25-minmax-neon.c326 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
328 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
447 const float32x4_t vk22x0123 = vld1q_f32(w + 180); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
448 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
557 const float32x4_t vk22x0123 = vld1q_f32(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon() local
558 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon()
Dup8x25-minmax-neonfma.c326 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
328 vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
447 const float32x4_t vk22x0123 = vld1q_f32(w + 180); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
448 vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
557 const float32x4_t vk22x0123 = vld1q_f32(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma() local
558 vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma()
Dup8x25-wasmsimd.c369 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
371 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
540 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
541 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
655 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd() local
656 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_ukernel_up8x25__wasmsimd()
Dup8x25-minmax-neonfma-acc2.c326 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
328 vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
450 const float32x4_t vk22x0123 = vld1q_f32(w + 180); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
451 vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
562 const float32x4_t vk22x0123 = vld1q_f32(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2() local
563 vacc0123p0 = vfmaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neonfma_acc2()
Dup8x25-minmax-wasmsimd-arm.c371 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
373 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
545 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
546 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
661 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm() local
662 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm()
Dup8x25-minmax-neon-acc2.c326 const float32x4_t vk22x0123 = vld1q_f32(w); w += 4; in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
328 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
450 const float32x4_t vk22x0123 = vld1q_f32(w + 180); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
451 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
562 const float32x4_t vk22x0123 = vld1q_f32(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2() local
563 vacc0123p0 = vmlaq_f32(vacc0123p0, vi22x0123, vk22x0123); in xnn_f32_dwconv_minmax_ukernel_up8x25__neon_acc2()
Dup8x25-minmax-wasmsimd-x86.c371 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
373 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
545 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
546 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
661 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86() local
662 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86()
Dup8x25-minmax-sse-acc2.c371 const __m128 vk22x0123 = _mm_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
373 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
547 const __m128 vk22x0123 = _mm_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
548 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
665 const __m128 vk22x0123 = _mm_load_ps(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2() local
666 vacc0123p0 = _mm_add_ps(vacc0123p0, _mm_mul_ps(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__sse_acc2()
Dup8x25-minmax-wasmsimd-x86-acc2.c371 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
373 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
548 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
549 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
666 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2() local
667 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_x86_acc2()
Dup8x25-minmax-wasmsimd-arm-acc2.c371 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
373 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
548 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
549 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()
666 const v128_t vk22x0123 = wasm_v128_load(w + 184); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2() local
667 vacc0123p0 = wasm_f32x4_add(vacc0123p0, wasm_f32x4_mul(vi22x0123, vk22x0123)); in xnn_f32_dwconv_minmax_ukernel_up8x25__wasmsimd_arm_acc2()